# Are we chasing ghosts? Quantifying unattributable polarization, and attributing the rest to annotator groups

arXiv: https://arxiv.org/abs/2602.06055 · HTML: https://mlsanity.com/paper/2602.06055/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 9 of its 27 references that are in the map.
- First posted on arXiv 16 Jan 2026.

## Builds on (most cited first, 9 of 9)

- [Annotators with Attitudes: How Annotator Beliefs And Identities Bias Toxic Language Detection](https://mlsanity.com/paper/2111.07997/) (2021)
- [DICES Dataset: Diversity in Conversational AI Evaluation for Safety](https://mlsanity.com/paper/2306.11247/) (2023)
- [Is Your Toxicity My Toxicity? Exploring the Impact of Rater Identity on Toxicity Annotation](https://mlsanity.com/paper/2205.00501/) (2022)
- [Designing Toxic Content Classification for a Diversity of Perspectives](https://mlsanity.com/paper/2106.04511/) (2021)
- [Like trainer, like bot? Inheritance of bias in algorithmic content moderation](https://mlsanity.com/paper/1707.01477/) (2017)
- [Cross-replication Reliability -- An Empirical Approach to Interpreting Inter-rater Reliability](https://mlsanity.com/paper/2106.07393/) (2021)
- [Human and LLM Biases in Hate Speech Annotations: A Socio-Demographic Analysis of Annotators and Targets](https://mlsanity.com/paper/2410.07991/) (2024)
- [NUTMEG: Separating Signal From Noise in Annotator Disagreement](https://mlsanity.com/paper/2507.18890/) (2025)
- [Mixture polarization in inter-rater agreement analysis: a Bayesian nonparametric index](https://mlsanity.com/paper/2309.15076/) (2023)

## Abstract (arXiv)

Standard agreement metrics often fail to capture systematic differences in opinion between minority and majority-group annotators, jeopardizing tasks such as hate speech and toxicity detection. Polarization has recently been proposed as a more robust way of distinguishing minor disagreements from systematic differences in opinion, but existing approaches do not provide practical tools for attributing it to specific annotator groups. We evaluate current methods and identify two major limitations in realistic settings: (1) the presence of ``inherent'' polarization that cannot be attributed to any known or latent groups, and (2) opposing polarization effects canceling each other out in aggregated annotations. To address these issues, we introduce a new metric that measures and tests the statistical significance of polarization attribution for annotator groups while avoiding these limitations, as well as an open-source Python library implementation, finding that no more than 20 annotators are needed per comment for reliable estimation. We apply our method to four subjective NLP datasets and find that gender and race consistently explain polarization patterns, while differences between annotator groups become stronger as the groups are further apart.